Jakub Jelinek wrote :
On Tue, Aug 31, 2004 at 05:38:31PM +0200, Matthias Saou wrote:
> Is then having the same library twice, the regular one in /usr/lib and
> the SSE2 optimized one in /usr/lib/sse2, expected to "just work" at
> runtime? If
Yes, it will just work.
Both the dynamic linker and ldconfig know how to handle it.
> so, I didn't know the existence of this, and will definitely look into
> it. What about MMX? Should one just simplify with SSE vs. non-SSE
> instead and put (non runtime) MMX optimized libs there too?
ATM sse2 is the only "important" feature ld.so on IA-32 handles.
Previously it used to be mmx, but as every added feature slows down
library loading when not using ld.so.cache (e.g. when LD_LIBRARY_PATH
is used or DT_RPATH; every feature doubles the number of stat'ed
directories before the non-existing directory cache is filled),
it was just changed to sse2 instead of adding sse2 to mmx.
SSE2 was chosen because you can get quite a big speedup already
by recompiling with -msse2 -mfpmath=sse.
Thanks for this valuable insight. I'll dig into a few relevant multimedia
packages and make a few "plain vs. optimized" tests of my own to see what
gives.
Matthias
--
Clean custom Red Hat Linux rpm packages :
http://freshrpms.net/
Fedora Core release 2 (Tettnang) - Linux kernel 2.6.8-1.521
Load : 1.00 1.08 1.11